MAGNN training memory consumption

[esprit1995]

Hello,

First of all, thank you for your paper and your code, it is a big help for anyone interested in HNE problem.
I have a question regarding MAGNN training. I tried fitting it to unattributed PubMed (smallest dataset of 4) in an unsupervised fashion. However, I couldn't do it - after the training started, all 32 gigs of RAM I have available were taken and then the script crashed.
I didn't change MAGNN parameters after cloning the repo and ran the data transform stage as indicated in readme. My question is, am I doing something wrong? Not entirely sure how 118 MB dataset and 2-layer net could do this. How much memory did you need for this task while obtaining the results for the paper?
I have seen in the closed issues that someone else has run into the same problem but there was no definite resolution there :(
Please help :)

[xiaoyuxin1002]

The data transform stage identifies all the 1-hop and 2-hop meta paths by default. Since there are 10 different link types in PubMed, there will be a relatively huge number of meta paths used for training. You may wish to do some selection on those meta paths before feeding them into MAGNN. I would suggest you start with a total of 10 different meta paths and see if that would solve your problem.

[esprit1995]

That helped, thanks. I left 1-hop metapaths to start with.
Could you share the metapaths you used to obtain the results presented in the paper, please?